sana

SANA is an efficiency-oriented open-source codebase from NVIDIA Labs for high-resolution image and video generation. It bundles complete training and inference pipelines for a family of diffusion transformer models, including SANA, SANA-1.5, SANA-Sprint, SANA-Video, SANA-Video 2.0, SANA-WM, SANA-Streaming and Sol-RL. What it provides: - Text-to-image generation up to 4K resolution, with model variants at 600M and 1.6B scale and multilingual (English, Chinese, emoji) fine-tuned checkpoints. - SANA-1.5: training-time and inference-time compute scaling for improved quality. - SANA-Sprint: one/few-step generation via sCM distillation, reported at 0.1s per 1024px image on H100. - SANA-Video and LongSANA: video generation using Block Causal Linear Attention, with text-to-video and text-image-to-video support. - SANA-Video 2.0: 5B and 14B architectures with hybrid linear/softmax attention and Attention Residuals; a 5B 720p checkpoint and a 4-step DMD preview are published. - SANA-WM: a 2.6B controllable world model for 720p, minute-scale video with 6-DoF camera control. - SANA-Streaming: a 2B real-time streaming video-to-video editing model for 720p, minute-scale video. - Sol-RL: reinforcement-learning post-training recipes using NVFP4 rollout and BF16 training, with recipes for SANA, FLUX.1 and SD3.5-L. Key techniques described in the repository include linear attention in the DiT backbone, DC-AE 32x image compression to reduce latent tokens, a decoder-only text encoder, Block Causal Linear Attention with Causal Mix-FFN for long video, Flow-DPM-Solver sampling, sCM distillation, and low-precision RL rollout. The project states that 4-bit quantization allows deployment on GPUs with under 8GB VRAM. Ecosystem integration: models are available through Hugging Face, diffusers pipelines, ComfyUI nodes, SGLang serving with an OpenAI-compatible API, Cosmos-RL post-training, and Replicate API. Demos are hosted for SANA, 4-bit, ControlNet, SANA-Sprint, SANA-Video 2.0, SANA-WM and SANA-Streaming. Getting started: clone the repository and run the provided environment setup script, then use the documented inference and training configs. The README includes a full example command for generating a 720p video with the SANA-Video 2.0 5B checkpoint. The codebase is licensed under Apache 2.0.